Invest1 publisher3 min readPublished
Salesforce runs Agentforce a generation behind the models its keynote guests debated
Nvidia's Jensen Huang told frontier labs to run as fast as they can. On the Dreamforce floor, the Salesforce customers and partners CNBC quoted said last year's models already cover the work they sell, and one of them would welcome a pause.
The Investor · Invest desk
What happened
- Nvidia CEO Jensen Huang, on stage with Marc Benioff at the 50,000-person Dreamforce conference in San Francisco, urged frontier labs to "run as fast as you can."
- Days earlier an Anthropic researcher resigned saying the top labs were "gambling with our lives", after which Dario Amodei and Sam Altman proposed safety initiatives and a slower pace of model development.
- Salesforce customers and partners at the event told CNBC that older and cheaper AI models are plenty powerful for everyday sales and customer service work.
- Attendees on the conference grounds said they are having trouble taking advantage of the AI technology that already exists, rather than worrying about how fast the next models arrive.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- constraint If most agentic output comes from last year's models, a lab charging a premium for the next generation has to find that demand somewhere other than packaged enterprise software.
- cost A development pause would be paid for by the labs in forgone upgrade revenue, not by buyers who say one generation back already covers their workloads.
- decision Buyers weighing Anthropic and OpenAI against cheaper open-source models are now deciding on price, because on the work they described the capability question is already settled.
- contradiction The same conference produced Huang's instruction to run as fast as possible and Nice's view that everything past today's models is icing, and the two are only reconcilable if model capability and deployable capability are priced separately.
CNBC's case for the absorption problem rests on four named people from the floor. Alec Bronston is a senior Salesforce director at the Chicago retail data company Spins [5]; Jaya Rohit Vuyyuru a vice president at the consultancy SummitX [14]; Tim Sanders chief innovation officer at the software review company G2 [16]; Kevin Lee technology chief at the contact centre software vendor Nice [17]. Four badges out of 50,000 is 0.008 percent [1]. Three of the four sell software or services into this ecosystem, and one runs Salesforce inside a customer [2].
The firmer item is a web page. For bots built with Agentforce, Salesforce is not relying on Anthropic's Claude Fable 5.1 or OpenAI's GPT-6 Astra, according to a support page on its own site [15]. Benioff has spent months arguing that AI is an accelerant to Salesforce's business and pushing back on the "SaaSpocalypse" narrative [8]. On the August earnings call the company unveiled Claudeforce, a way for salespeople to reach critical data from inside Anthropic's Claude chatbot [10].
"The majority of agentic outcomes aren't driven by frontier capabilities," Sanders said. "They're driven by last year's AI." [16][18] Lee said his customers are served by "the models that are out there already today, and even one generation behind". Everything beyond that point, he said, is "icing on the cake" [17]. Huang, on stage with Benioff, urged the labs to "run as fast as you can" [2]. Bronston's problem is a different one: "It's already hard enough to keep up," he said, and a slowdown would give him "a lot of opportunity to just even catch up and get our feet wet" [5].
Many executives told CNBC that AI has been transformational for their business [19]. Anthropic's booth was the most trafficked on the exposition floor on Tuesday, staffed by employees in oversized white sweaters running demonstrations for passersby [11]. A preference for last year's model may be a procurement lag; many users are still working out their AI budgets and weighing Anthropic and OpenAI against cheaper open-source alternatives [13].
In my view the gap holds through at least one more release cycle, because a better model does not help a buyer whose constraint is keeping up with the one he already has [5]. The test is narrow and public. If the Agentforce support page starts listing Fable 5.1 or GPT-6 Astra for production bots, the absorption reading is wrong [15]. Salesforce shares are down 8% this year even after a large jump in August, while Adobe and Autodesk are down significantly more [9]. That is a price on software vendors. It is not a measure of what anyone will pay for the next model.
What to watch
- Whether the slower pace Amodei and Altman proposed changes actual release timing at Anthropic and OpenAI.
- Whether other agent platforms follow Salesforce in publishing which model generation their production bots call.